Custom-SVG – An SVG custom generation framework jointly launched by Adobe and the City University of Hong Kong

What is Custom-SVG

Custom-SVG is a two-stage, style-customizable SVG generation framework developed by Adobe and City University of Hong Kong. It generates high-quality SVG graphics with personalized styles based on text prompts. The framework introduces a two-stage pipeline where a path-level diffusion model learns the structure and content of SVGs, while a customized text-to-image (T2I) model provides style information, enabling diverse style customization.
By combining the efficiency of feedforward models with the powerful generative capabilities of diffusion models, Custom-SVG quickly produces well-structured and stylistically consistent vector graphics, making it ideal for use in design, icon creation, and related applications.

Key Features of Custom-SVG

• Maintains Structured SVG Layouts: Generates vector graphics with clear hierarchical structures and regular paths, making them easy to edit.

• Style Customization: Produces SVGs in diverse styles that match a small set of user-provided style examples.

• Efficient Generation: Utilizes a feedforward-based approach to ensure fast generation, suitable for practical design workflows.

• Semantic Alignment: Ensures that the generated SVG content accurately aligns with the input text prompts.

Technical Overview of Custom-SVG

• T2V Diffusion Model Based on Path-Level Representation:
  ◦ Path-Level Representation: SVGs are decomposed into path-level representations, with each path defined by Bézier curves, color, and other parameters. These are encoded into compact latent vectors using a pretrained SVG VAE.
  ◦ Diffusion Model: A denoising process recovers SVG tensors aligned with text prompts, gradually refining output from noise.
  ◦ Transformer Architecture: A Transformer backbone with self-attention and cross-attention mechanisms helps the model understand textual semantics and generate matching SVG structures.

• Style Customization via Image Diffusion Prior:
  ◦ Style Extraction: A T2I diffusion model is fine-tuned on a few style examples to generate diverse stylized images.
  ◦ Image-Level Loss: Generated SVGs are rendered into images and optimized using image-level loss to guide the T2V model toward the desired visual style.
  ◦ Style Transfer: Style information from the customized images is transferred to the SVG generation process, enabling flexible visual styling.

Project Links for Custom-SVG

• Official Website: https://customsvg.github.io/

• GitHub Repository: https://github.com/intchous/custom-svg-style

• arXiv Paper: https://arxiv.org/pdf/2505.10558

Application Scenarios for Custom-SVG

• Graphic Design & Illustration: Quickly generate vector graphics in specific styles for concept visualization, illustration, and brand design.

• User Interface (UI) Design: Create icons, buttons, and other UI elements tailored to different design themes.

• Web Design: Produce dynamic vector graphics to enhance visual appeal and responsive design of websites.

• Education & Training: Generate instructional graphics and interactive learning tools to support teaching.

• Personalized Content Creation: Generate custom vector graphics for personalized gifts, social media content, and more.